Issues

• FP8 dense: per-tensor activation scale couples the sequences in a batch

  gdevenyi/FreeToken#94 · Filter by author gdevenyi opened on Oct 20, 2026

• qwen3_coder bare <function=...> (no <tool_call>) is a call in non-streaming and plain text in streaming

  gdevenyi/FreeToken#90 · Filter by author gdevenyi opened on Oct 15, 2026
  1

• qwen3_coder non-streaming drops the text between tool calls

  gdevenyi/FreeToken#89 · Filter by author gdevenyi opened on Oct 15, 2026
  1

• qwen3 whitespace around </think> differs between streaming and non-streaming

  gdevenyi/FreeToken#88 · Filter by author gdevenyi opened on Oct 15, 2026
  1

• qwen3 reasoning parser re-enters reasoning on a later <think> and deletes the markers

  gdevenyi/FreeToken#87 · Filter by author gdevenyi opened on Oct 15, 2026
  1

• Cold greedy output occasionally differs between server starts under the new decode numerics (upstream #631/#585), think prompt only

  gdevenyi/FreeToken#83 · Filter by author gdevenyi opened on Oct 8, 2026

• transformers 5.15.1 AutoTokenizer drops \p{M} from Qwen3.8's pre-tokenizer regex; combining-mark text tokenizes differently from tokenizer.json

  gdevenyi/FreeToken#80 · Filter by author gdevenyi opened on Oct 7, 2026
  1

• Merging upstream FlashML-org/FreeToken#588 (d3512b4) as-is deadlocks the filler-thread PLE graph fill (fork #67) at decode step 2

  gdevenyi/FreeToken#78 · Filter by author gdevenyi opened on Oct 2, 2026

• research: speculative decoding for qwen4_exp on the 5080 (investigated, low expected value; hand-off for a future agent)

  gdevenyi/FreeToken#70 · Filter by author gdevenyi opened on Sep 26, 2026

• decode deadlock: PLE disk wait-sync sets its flag only after cuGraphLaunch returns, which can block (hybrid MoE graphs)

  gdevenyi/FreeToken#65 · Filter by author gdevenyi opened on Sep 24, 2026
  4

• runtime cache rebuild accepts a geometry that OOMs on the next prefill chunk

  gdevenyi/FreeToken#66 · Filter by author gdevenyi opened on Sep 24, 2026

• FREETOKEN_FP8_DENSE=1 with a bf16 FTW checkpoint dies with KeyError in_proj_qkvz.weight

  gdevenyi/FreeToken#8 · Filter by author gdevenyi opened on Sep 15, 2026

• --moe-collect-stats: oracle_hit is a static top-C bound, not an upper bound, and LRU beats it

  gdevenyi/FreeToken#7 · Filter by author gdevenyi opened on Sep 15, 2026

• Intermittent '</think>' in message.content on a tool call (qwen3 reasoning parser, 1 of 13)

  gdevenyi/FreeToken#6 · Filter by author gdevenyi opened on Sep 14, 2026
  1

• ft ctl cache: growing the MoE pool while shrinking KV OOMs and rolls back on a nearly full card

  gdevenyi/FreeToken#5 · Filter by author gdevenyi opened on Sep 14, 2026

• --moe-cache-auto fails only after the expert-bank load, with a 1024-slot floor that ignores the byte budget

  gdevenyi/FreeToken#4 · Filter by author gdevenyi opened on Sep 14, 2026

• GLM DSA sparse kernel exceeds sm_120 Triton shared-memory limit after merging #408

  gdevenyi/FreeToken#2 · Filter by author gdevenyi opened on Sep 14, 2026

• Qwen3.8-Flash-Next: GDN prefill OOM on a long prompt kills the backend (prefill transient not budgeted)

  gdevenyi/FreeToken#3 · Filter by author gdevenyi opened on Sep 14, 2026
  1